henhaobio
Sign in

Proceedings of the National Academy of SciencesGenomicsPeer-reviewed journal article

RepeatModeler2: automated genomic discovery of transposable element families

Jullien M. FlynniD, Robert M. HubleyiD, Clément GoubertiD, Jeb R. RoseniD, Andrew G. ClarkiD, Cédric FeschotteiD, Arian F. A. SmitiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

The accelerating pace of genome sequencing throughout the tree of life is driving the need for improved unsupervised annotation of genome components such as transposable elements (TEs). Because the types and sequences of TEs are highly variable across species, automated TE discovery and annotation are challenging and time-consuming tasks. A critical first step is the de novo identification and accurate compilation of sequence models representing all the unique TE families dispersed in the genome. Here we introduce RepeatModeler2, a new pipeline that greatly facilitates this process. This new program brings substantial improvements over the original version of RepeatModeler, one of the most widely used tools for TE discovery. In particular, this version incorporates a module for structural discovery of complete LTR retroelements, which are widespread in eukaryotic genomes but recalcitrant to automated identification because of their size and sequence complexity. We benchmarked RepeatModeler2 on three model species with diverse TE landscapes and high-quality, manually curated TE libraries: Drosophila melanogaster (fruit fly), Danio rerio (zebrafish), and Oryza sativa (rice). In these three species, RepeatModeler2 identified approximately three times more consensus sequences matching with >95% sequence identity and sequence coverage to the manually curated sequences than the original RepeatModeler. As expected, the greatest improvement is for LTR retroelements. The program had an extremely low false positive rate when applied to simulated genomes devoid of TEs. Thus, RepeatModeler2 represents a valuable addition to the genome annotation toolkit that will enhance the identification and study of TEs in eukaryotic genome sequences. RepeatModeler2 is available as source code or a containerized package under an open license (https://github.com/Dfam-consortium/RepeatModeler, https://github.com/Dfam-consortium/TETools). SignificanceGenome sequences are being produced for more and more eukaryotic species. The bulk of these genomes is composed of parasitic, self-mobilizing transposable elements (TEs) that play important roles in organismal evolution. Thus there is a pressing need for developing software that can accurately identify the diverse set of TEs dispersed in genome sequences. Here we introduce RepeatModeler2, an easy-to-use package for the curation of reference TE libraries which can be applied to any eukaryotic species. Through several major improvements over the previous version, RepeatModeler2 is able to produce libraries that recapitulate the known composition of three model species with some of the most complex TE landscapes. Thus RepeatModeler2 will greatly enhance the discovery and annotation of TEs in genome sequences.

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 1 (26 Nov 2019) · CC BY 4.0

With a new structural LTR module, RepeatModeler2 recovers 2.9 to 4.7 times more curated TE families near-perfectly than RepeatModeler 1.0.11 in fruit fly, zebrafish, and rice.

Repeat masking precedes gene annotation, yet transposable element (TE) family discovery in new genomes remains largely manual, and automated tools, including the original RepeatModeler, emit fragmented, partly redundant consensi that inflate family counts, a problem acute for long terminal repeat (LTR) retroelements. RepeatModeler2 adds structural LTR discovery plus cross-module redundancy reduction and is benchmarked against curated libraries of three model species.

Key results

  1. ShownRepeatModeler2 (RM2) reconstructs many more curated TE families as a single high-identity, near-full-length consensus than RepeatModeler 1.0.11 (RM1) in all three species, mostly through LTR families RM1 missed.

    'Perfect' = one library sequence matching a curated consensus at >95% nucleotide similarity and >95% length coverage; RM2 (-LTRStruct) gave 2.9 to 4.7 times more perfect families than RM1 (defaults) versus Repbase (fly, zebrafish) and a curated rice library, plus more close matches overall; gain mostly from LTR families. · Results and Discussion; Benchmarking; Benchmarking Parameters; Figure legends

  2. ShownMasking each genome with its RM2 library reproduces the subclass-level TE landscape of the curated libraries: LTR-rich, retrotransposon-dominated fruit fly; terminal inverted repeat (TIR) DNA transposon-dominated zebrafish with diverse LTRs; rice with near-equal TIR and LTR shares.

    RepeatMasker/parseRM genome fractions and per-subclass family counts (Figure 2) track the curated profiles (no numeric values in text); in zebrafish RM2 defines about half as many LTR families and twice as many TIR families as Repbase, attributed to looser reference LTR family definitions and TIR fragmentation. · Results and Discussion; Benchmarking; Figure legends

  3. ShownRM2 library sizes are comparable to the curated references, and despite an added independent LTR module they are not drastically larger than RM1 output, owing to CD-HIT merging that keeps LTR-module representatives and tags near-duplicate RepeatScout/RECON families as putative subfamilies.

    RM2 output 766 (D. melanogaster), 3851 (D. rerio), and 2648 (O. sativa) library sequences, which the authors call comparable to curated library counts (Table 1); families clustering within 20% similarity are labeled putative subfamilies, and counts excluding them are also reported. · Results and Discussion; Combining libraries and reducing redundancy; Tables

  4. SuggestedOn transposon-free synthetic genomes, RM2 reports almost no spurious families and its LTR structural module reports none, indicating a low false-positive rate.

    GARLIC-simulated TE-free genomes mimicking each genome's complexity, isochores, and tandem repeats: one false-positive family for the D. melanogaster model, five for the D. rerio model, and none from the LTR module. · Results and Discussion

  5. SuggestedRM2 also captures element classes traditionally viewed as hard for ab initio discovery, namely MITEs (miniature inverted-repeat transposable elements, typically non-coding) and Helitrons, without apparent shortfall versus curated references.

    Based on Figure 2 subclass profiles: zebrafish and rice, both rich in MITE families, showed no MITE under-representation in RM2 libraries, and Helitron identification compared well with the references; the text gives no per-class sensitivity numbers. · Results and Discussion

Methods, data and reproducibility

RepeatScout analyzes a random 40 Mbp sample, then RECON rounds run on pre-masked samples tripling from 3 Mbp (default cap 243 Mbp or 5 rounds); an optional LTRharvest/LTR_retriever module with MAFFT/Ninja clustering adds LTR families, Refiner builds seed alignments and CpG-aware consensi, CD-HIT merges modules, and RepeatClassifier assigns classes.

  • D. melanogaster PBcr-BLASR Celera 8.1 genome assembly (Koren et al. 2012), input for TE discovery and masking benchmarks · D. melanogaster (fruit fly)
  • D. rerio danRer10 genome assembly, input for TE discovery and masking benchmarks · D. rerio (zebrafish) · GCF_000002035.5
  • O. sativa version 7.0 pseudomolecule genome assembly, input for TE discovery and masking benchmarks · O. sativa (rice)
  • Repbase curated TE consensus library for D. melanogaster (release 20181026), reference truth set · D. melanogaster (fruit fly)
  • Repbase curated TE consensus library for D. rerio (release 14.01), reference truth set · D. rerio (zebrafish)
  • Manually improved O. sativa TE library from a cited earlier study, reference truth set · O. sativa (rice)
  • GARLIC-simulated TE-free artificial genomes modeled on D. melanogaster and D. rerio, for false-positive testing · none (synthetic; modeled on D. melanogaster and D. rerio)
Methods & tools
RepeatModeler2 (Refiner, RepeatClassifier with RepeatMasker Repeat Protein DB and Dfam/RepBase), RepeatModeler 1.0.11 (baseline), RepeatScout 1.0.6, RECON 1.0.8, LTRharvest, LTR_retriever 2.6, MAFFT 7.407, Ninja 0.97-cluster_only, CD-HIT 4.8.1, RepeatMasker 4.0.9-p2, RMBLast 2.9.0-p2, TRF 4.0.9, parseRM.pl, GARLIC (TE-free artificial genome simulator)
Software & environment
CentOS 7.6.1810 Linux; Docker/Singularity containers available; dependencies: CD-HIT 4.8.1, GenomeTools 1.5.10, LTR_retriever 2.6, MAFFT 7.407, Ninja 0.97-cluster_only, RECON 1.0.8, RepeatMasker 4.0.9-p2, RepeatScout 1.0.6, RMBLast 2.9.0-p2, TRF 4.0.9.
Assay conditions & compute
Shared research-group Linux server with an Intel Xeon E7-4830 v4 (2.00GHz) CPU, 112 cores, and 504 GB RAM; pipeline targets a single multi-processor computer; per-run runtimes and -srand seeds are tabulated only in Table S2.
Reproducibility gaps
No download location is stated for the RM2 benchmark libraries; the D. melanogaster PBcr-BLASR Celera 8.1 assembly and the manually improved rice library are cited without URLs or accessions; Repbase references are given by release only.
Limitations (6)
  • RepeatClassifier is rudimentary: it relies solely on similarity to catalogued TEs and TE protein domains, hampering classification of non-coding or highly diverged families; structural cues such as TIR motifs or target site duplications are not yet used. · stated by the authors
  • DNA-TIR families are discovered only by the RepeatScout/RECON rounds, which yield shorter, more fragmented consensi than the LTR module, so TIR family counts (about twice the reference in zebrafish) may be inflated by variants or fragments of single families. · stated by the authors
  • LTR_retriever does not resolve every nested insertion, which can degrade the LTR library; the optional -LTRMaxSeqLen cap (10000 used for D. melanogaster) mitigates this but may bias results, so authors advise it only when maximum LTR length is known or nesting is problematic. · stated by the authors
  • Family-level scoring thresholds (>95% or >80% identity and coverage) are strict and likely account for the many reference families scored 'not found', even where fragments may exist and genome-wide subclass masking broadly matches the curated libraries. · stated by the authors
  • Benchmarks compare RM2 only with RepeatModeler 1.0.11 on three well-curated model genomes, apparently one seeded run each; no other de novo TE discovery pipeline is tested and run-to-run variation from random genome sampling is not quantified. · noted in this summary
  • False-positive testing used only two simulated TE-free genomes (fly and zebrafish models); rice was not simulated, and synthetic backgrounds may not capture real-genome confounders such as multicopy gene families or segmental duplications. · noted in this summary

Read this if you build repeat libraries for newly assembled eukaryotic genomes and want to see how RepeatModeler2, with its structural LTR module, performs against RepeatModeler 1.0.11 and curated fly, zebrafish, and rice libraries.

Citations
3,844
FWCI
201.72
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISRodriguez '21Zeng '18Qian '25Orozco-Aria '23Bickmann '25Hu '23
76% semantic matchBioinformatics
Matías Rodriguez, Wojciech Makałowski·bioRxiv·Feb 8, 2021· 1 citations